UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolution of Image Generation and Detection

Forum topic · 小凯 · 2026-04-27

Summary

UniGenDet is a unified generative-discriminative framework that enables image generation and generated-image detection to co-evolve, addressing the fact that the two fields have developed largely independently with distinct architectural paradigms. Proposed by Yanran Zhang, Wenzhao Zheng, Yifei Li, Bingyao Yu, Yu Zheng, Lei Chen, Jiwen Lu, and Jie Zhou (arXiv:2604.21904), the framework bridges the task gap via a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm. Through this synergy, the generation task improves the interpretability of authenticity recognition, while authenticity criteria guide the creation of higher-fidelity images. The authors further introduce a detector-informed generation alignment mechanism to facilitate seamless information exchange between the two tasks. Extensive experiments on multiple datasets demonstrate state-of-the-art performance, showing the potential of adversarial information to jointly enhance generative quality and detection robustness.

Overview

Research area: Computer Vision (CV) Authors: Yanran Zhang, Wenzhao Zheng, Yifei Li, Bingyao Yu, Yu Zheng, Lei Chen, Jiwen Lu, Jie Zhou Published: 2026-04-23 arXiv: 2604.21904

Abstract (Translation)

In recent years, both image generation and generated-image detection have made significant progress. Despite their rapid development, they have evolved largely independently, and the two fields have formed distinct architectural paradigms: the former relies mainly on generative networks, while the latter favors discriminative frameworks. A recent trend in both fields is the use of adversarial information to improve performance, revealing their potential for synergy. However, the substantial architectural differences between them pose considerable challenges.

Unlike previous approaches, the authors propose UniGenDet: a unified generative-discriminative framework for the co-evolution of image generation and generated-image detection.

Key Contributions

Symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm to bridge the task gap between generation and detection.
Through this synergy, the generation task improves the interpretability of authenticity recognition, while authenticity criteria guide the creation of higher-fidelity images.
A detector-informed generation alignment mechanism is introduced to facilitate seamless information exchange between the two tasks.
Extensive experiments on multiple datasets show that the method achieves state-of-the-art performance.